Learning to shine: Neuroevolution enables optical control of phase transitions

Abstract

We address the problem of active optical steering of structural phase transitions in solids. We demonstrate that existing reinforcement learning approaches can derive optimal time-dependent electric fields in optically-driven dissipative classical systems far beyond the harmonic regime, enabling the stabilization of non-thermal structural phases. Our approach relies on experimentally extractable metrics of the phase-space evolution and physically-interpretable Fourier Neural Network surrogates of the time-dependent electric field. Using first-principles simulations, we demonstrate the stabilization of a symmetric phase in bismuth through impulsive Raman scattering under continuous and pulsed light sources in the presence of dissipation. Importantly, the method is gradient-free, which enables optimization loops based solely on experimental data, such as the measures of half-periods of oscillations in transient spectroscopy. Our framework thus provides a practical route for controlling non-equilibrium structural dynamics with light, opening pathways to stabilize hidden and metastable phases in quantum materials.

Figures (22)

• Figure 1: (a) Potential energy landscape along the A$_{1g}$ displacement vector (top) and the corresponding Raman cross-section (bottom), (b) Time evolution of phonon amplitude starting from three different initial conditions, as marked in (a). Waterfall plot of the Fourier transforms of the oscillations from small (top) to large (bottom) amplitude oscillations. (c) and (d) Phase portraits of the system evolution under sinusoidal (c) and pulsed (d) harmonic protocols.
• Figure 2: (a) Optimization framework for time-varying amplitude of electric field using a Fourier neural network (FNN) architecture. The time-dependent output function is a continuously varying amplitude (shown in right is amplitude function for four different initial NN parameters) (b) Score of each protocol in each generation, (c) phase space portraits (left) corresponding to the topmost protocols (right) for the three selected generations, (d) Fourier transformation of the respective time dependent protocols in (c).
• Figure 3: (a) An example of a pulse driven protocol obtained using the same FNN architecture (detailed description is in Figure \ref{['fig:ifft_pulse']} of the SM), (b) Score of best performing protocol in each generation, (c) phase space portraits (left) corresponding to the topmost pulsed protocols (right) for the three selected generations, (d) distribution of protocol efficiency ($\eta$) and pulse timing spread ($\Delta t$) across generations (Only top ten from each generation are plotted for clarity; for detailed plots, see Figure \ref{['fig:eff_groups']} of the SM).The red inverted triangle indicates the optimal protocol with a score of 1
• Figure 4: (a) Illustration of a two-pulsed protocol (top panel) and the corresponding time evolution of position (bottom), (b) Harmonic pulse driven protocols with varying time delays (top) and the corresponding normalized imaginary part of the dielectric function with time (bottom), (c) Correlation of the ratio (as defined in (a)) with the average position for all harmonic protocols; eliminating the data in red and blue (refers to the points in the time window around the pulse timing as marked in (a)) gives rise to linear relationship between the two quantities (bottom panel)
• Figure S1: (a) Full–basis CW optimized protocol using all ten sinusoidal components and, (b) Five-component subset that zeros five output weights but attains essentially the same score as (a)